Abstract

We identified, trained, and optimized a machine learning model to identify vehicles based on make, model, and year to be used by Hunter Engineering in conjunction with their QuickID system. QuickID is a mounted camera and software system that Hunter Engineering uses to identify moving cars prior to servicing. There is currently a very limited data set with only a few classification parameters, which prevents us from building a robust deep learning model from scratch. Our project uses a method called transfer learning, which allows us to apply knowledge from existing deep learning models to solve the classification problem. We used the open source toolkit TensorFlow for Python to retrain, optimize and test the robust Inception-v3 model. We designed the system to improve upon QuickID’s accuracy (situationally limited to 92%), with a goal confidence level of > 40%. We first optimized three hyperparameters – number of steps, training batch size, and learning rate – assessing these parameters by validation accuracy and cross-entropy. Additionally, we created a decision rule that further increased accuracy and decreased the false positive rate of tested images outside the classes our model was trained on, i.e. data unknown to the model. Our final system has a false positive rate to 15% and identifies 70.15% of data with an accuracy of 98.85%  The other 29.85% of the data is classified as “unknown”. Overall, the project gives Hunter Engineering an idea of how feasible such a program would be as a validation method for QuickID. In the future, accuracy could be increased by better data characterization and optimizing additional hyperparameters.

Read the full report here.

Team

  • Mitali Avadhani
  • Mani Raman
  • Caroline Trier

Created in partial fulfillment of the requirement for the BS in Electrical Engineering and BS in Systems Science, School of Engineering and Applied Science, Washington University in St. Louis

Spring 2018